{"id":"6d3c586e-4514-48dc-9780-5651a21edc6f","arxiv_id":"2505.18784","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A physics-constrained smoothing method removes non-physical negative strains from DIC data and improves downstream learning of material constitutive laws on glove and heart-valve datasets.","lead":"Measurements of how materials deform in biaxial stretching tests are noisy and sometimes contain impossible negative strains. This paper adds a physics constraint to the data-smoothing step, producing cleaner strain fields that make learned material models of tissues more accurate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation is self-referential: PGS-smoothed fields serve as both training labels and test ground truth, so reported downstream error reductions do not establish that PGS extracts true deformation; synthetic ground-truth validation is required.","rationale":"The paper does several things well: the PGS formulation is clear, the number of introduced parameters is small (beta, RK support size), and the STest subset is an attempt to address non-physical data. However, the central claim is that PGS improves fidelity to the true deformation, and the only quantitative evidence is downstream task error measured on the processed data itself. This makes the evaluation self-referential: the same PGS-smoothed fields are used as both training targets and test ground truth, so lower test error for PGS-trained models may simply reflect that the learned operator imitates the PGS smoothing operator rather than recovering physical deformation. The STest subset is ambiguous because its ground truth is not documented, and no synthetic or independently known deformation field is used for validation. I partially agree with the reader's weakest assumption about the positivity constraint: forcing dominant-direction normal strains to be non-negative could remove true local compressive strains arising from heterogeneity, shear, or anisotropic Poisson effects, and this deserves scrutiny. But even if the positivity constraint is physically correct, the paper's evidence would still be inconclusive due to the self-referential evaluation. Therefore, the verdict should remain CONDITIONAL: the method is plausible and clearly described, but the central claim requires validation against ground-truth deformation fields before it can be accepted.","tokens_in":12464,"tokens_out":7300,"duration_ms":67539,"concrete_test":"Build a synthetic biaxial DIC benchmark: simulate a known heterogeneous anisotropic hyperelastic material (e.g., TVAL-like with prescribed fiber orientation) under the same seven displacement-controlled protocols; generate ground-truth displacement/strain fields; add realistic sensor noise and smooth operation-error artifacts (e.g., negative strain bands in the dominant direction); run PGS, rough, and RK smoothing; compare reconstructed fields against ground truth via pointwise L2 and principal-strain error. Then train PNO on each dataset and test on a held-out set of ground-truth samples (not smoothed by PGS). If PGS improves field accuracy and downstream generalization relative to baselines, the central claim is supported; if not, the paper's empirical evidence is insufficient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PGS 'extracts' more accurate displacement/strain fields and 'induces smaller errors' in downstream learning is not supported by the reported metrics, because the evaluation is self-referential. In Tables 3 and 4, the relative L2 displacement error is computed against the same PGS-smoothed fields used as training labels; a model trained on PGS data will naturally fit PGS-smoothed targets better than rough or RK-smoothed targets. The STest subset (Table 4) is intended to break this circularity, but the ground truth for STest is not specified: if it is the PGS-processed displacement, the comparison is still circular; if it is the original noisy DIC measurement, then a lower error against noise is not evidence of physical accuracy. No synthetic or experimentally known deformation field is used to validate that PGS recovers the true field, and no error bars or code/data are provided. The positivity-constraint concern (Section 3, Eq. 9) is real but secondary: even if the sign rule is correct, the empirical demonstration would still be inconclusive without a ground-truth benchmark.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-guided smoothing (PGS) method for digital image correlation (DIC) displacement measurements. The method fits a reproducing-kernel displacement field to the raw DIC data while adding a soft penalty that suppresses negative normal strains (Section 3, Eq. (9)-(10)). The smoothed fields are then used as training labels for a peridynamic neural operator that learns a constitutive law and a fiber orientation field (Section 3, Eqs. (12), (22)-(25)). The authors evaluate PGS on two experimental datasets—a nitrile glove (Dataset 1) and a porcine tricuspid valve anterior leaflet (Dataset 2)—and report that downstream displacement-prediction errors are lower when the operator is trained on PGS-processed data than on raw or RK-smoothed data (Tables 3 and 4).","tokens_in":12687,"tokens_out":6394,"duration_ms":55171,"significance":"If the central claim is established, PGS would be a useful pre-processing tool for DIC-based material modeling, because it combines kernel smoothing with targeted physical constraints and is designed to feed into a downstream operator-learning pipeline. The paper includes useful ablations (Table 2), resolution generalization experiments (Table 3), and interpretable kernel/fiber-orientation visualizations (Figs. 3 and 5). However, the current evaluation does not validate that PGS recovers the true deformation field: the reported test errors are computed against the same pre-processed field type used as training labels, and the only independent physical anchor is a scalar average stress term in the training loss. The study lacks a synthetic ground-truth benchmark, which is needed to support the claim that PGS extracts physically accurate displacement and strain fields.","major_comments":[{"comment":"The evaluation is self-referential: for each data type, the relative L2 displacement error is computed against the same type of processed field used as training labels. A model trained on PGS-smoothed fields will naturally fit PGS-smoothed targets better than a model trained on raw or RK-smoothed fields fits its own targets, so the reported reductions (10.01% to 7.35% on Dataset 1; 6.21% to 5.11% on Dataset 2) do not establish that PGS extracts the true deformation. The paper does not specify the ground truth for the STest subset in Table 4; if STest labels are PGS-processed, the comparison remains circular, and if they are the original noisy measurements, a lower error against noise is not evidence of physical accuracy. The authors should validate PGS on synthetic data with known deformation fields (e.g., from finite element simulations with added sensor noise and operator artifacts) and report displacement/strain errors against the true field.","section":"Section 4, Tables 3 and 4"},{"comment":"The positivity constraint encodes a strong physical assumption: that negative normal strains, or negative strains in the direction of dominant tension, are always non-physical artifacts. The paper itself acknowledges in Section 2 that physical compressive strains can occur through Poisson contraction when applied tensions differ (references [22,23]). The protocol-specific rule in Section 4.2—penalizing E11 only for the first four loadings and E22 only for the last three—may be reasonable for a homogeneous, orthotropic specimen, but it is not justified for a heterogeneous biological tissue where local deformation can deviate from the global loading direction. If this assumption is violated for any sample, PGS will remove genuine compressive deformation and bias the training labels. The authors should test the sign rule on synthetic or known-field examples that include physically compressive zones, or provide direct experimental evidence that the penalized negative-strain regions are indeed artifacts.","section":"Section 3, Eq. (9) and Section 4.2"},{"comment":"No error bars, confidence intervals, or repeated random seeds are reported. This is particularly important for Dataset 2, where the overall test improvement of PGS over RK smoothing is marginal (5.13% vs. 5.11%, Table 4) and may not be statistically significant. Since hyperparameters (penalty weight, kernel support size) are selected on the same data, the authors should report mean and standard deviation over multiple training runs and data splits to demonstrate that the reported gains are robust.","section":"Tables 1-4 (general reporting)"}],"minor_comments":[{"comment":"Step 3 says 'using the analytical expression in (9)', but Eq. (9) is the physics-based loss; the reference should be to the least-squares solution in Eq. (6) or Eq. (11).","section":"Algorithm 1, step 3 and Eq. (11)"},{"comment":"The expression |ReLU(-E11) + ReLU(-E22)|^2 is redundant because both terms are nonnegative; if the intention was to penalize each component independently, the loss should be written as ReLU(-E11)^2 + ReLU(-E22)^2.","section":"Eq. (9)"},{"comment":"The manuscript contains numerous missing spaces and corrupted mathematical symbols (e.g., the abstract reads 'Inthiswork,wepresentanovelapproachtoprocesstheDIC'). A thorough proofread and cleanup of the LaTeX source is needed.","section":"General presentation"},{"comment":"The captions do not state that the reported values are relative L2 displacement errors; adding this to the captions would improve clarity.","section":"Tables 1 and 2"},{"comment":"The claim that PGS 'successfully eliminated negative strains' is illustrated on a representative sample; a quantitative summary (e.g., the percentage of nodes with negative strain before and after PGS across all samples) would make the claim more precise.","section":"Section 4.1, Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The central claim of the paper is not yet supported because the empirical evaluation is circular with respect to the processed fields. A synthetic ground-truth validation is essential and is a fixable issue within the manuscript's scope. The paper also does not provide code or data, which would be important for reproducibility. The manuscript's scope fits the journal, but the evaluation needs to be substantially strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The method is straightforward to grasp and the paper is honest about its design: RK-based smoothing with a ReLU penalty on negative Green-Lagrange strains, applied as a pre-processing step for downstream constitutive-operator learning. That specific recipe is new, and the ablation over the penalty weight and kernel support is useful. The two real DIC datasets, including the heterogeneous valve tissue, give the paper a practical feel. If the method works as advertised, it would be a cheap, useful fix for a real problem in biomechanics data pipelines.\n\nThe soft spot is the evaluation. The headline numbers in Tables 3 and 4 compare downstream errors against PGS-smoothed fields as ground truth — the same kind of field used to train the operator. So a model trained on PGS data naturally fits PGS-shaped targets better than rough or RK-smoothed ones. That is circular. The STest subset on Dataset 2 is meant to fix this, but the ground truth there is not specified. If STest labels are also PGS-processed, the circularity remains. If they are the original noisy DIC fields, then a lower error against noise is not evidence of physical accuracy. Without a synthetic benchmark where the true deformation is known, the central claim that PGS 'extracts' more accurate fields is not actually supported.\n\nThe positivity constraint itself is a reasonable assumption for these biaxial protocols, and the paper does acknowledge that true compressive strains can appear under unequal tension. But the protocol-specific sign rule (only E11 for some, only E22 for others) is asserted, not validated. That is a secondary concern after the circularity. The lack of error bars is also notable because the Dataset 2 gain on the overall test set (5.11% vs 5.13%) is within noise. No code or data are provided, which makes it harder to check.\n\nWho is this for? DIC users and researchers working on data-driven constitutive models. The idea deserves a serious look, but the empirical case needs a ground-truth benchmark and error bars. I would send it to peer review, but with the expectation that the authors add synthetic tests with known deformation fields and clarify what STest actually measures.","headline":"The PGS smoothing idea is sensible and clearly presented, but the evaluation does not establish that it recovers true deformation fields because the test labels are the same smoothed fields used for training.","tokens_in":13218,"tokens_out":1370,"would_cite":false,"duration_ms":13910,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adding a positivity constraint on normal strains during reproducing-kernel smoothing of digital image correlation measurements yields physically consistent displacement and strain fields and improves downstream…","keywords":["digital image correlation","physics-guided smoothing","reproducing kernel","strain field denoising","constitutive operator learning","peridynamic neural operator","biaxial stretching","biological tissue mechanics"],"falsifier":"Run PGS on a synthetic biaxial dataset whose ground-truth deformation is known and includes genuine compressive strains (e.g., from Poisson contraction under unequal tension), then compare the smoothed fields with the ground truth. If PGS systematically erases the true compressive regions or worsens the downstream learned operator compared with unconstrained smoothing, the sign rule is removing real physics rather than artifacts.","tokens_in":12258,"feed_emoji":"🔬","tokens_out":7085,"duration_ms":53095,"temperature":0.7,"pith_summary":"Digital image correlation (DIC) records full-field displacements of a specimen during biaxial stretching, but the measured fields contain sensor noise and smooth non-physical patterns from operating errors, and these artifacts corrupt strain fields used to learn material behavior. This paper introduces physics-guided smoothing (PGS), an optimization step that reconstructs the displacement field with reproducing-kernel basis functions subject to a penalty on negative normal strains. The paper aims to show that enforcing this partial physical knowledge at the data-processing stage produces displacement and strain fields that are both more physical and more useful for downstream tasks, concretely for learning a nonlocal constitutive law and the underlying fiber microstructure of biological tissue. On two real DIC datasets, a nitrile glove and a porcine tricuspid-valve anterior leaflet, PGS lowers the test relative $L^2$ displacement error of the learned constitutive operator compared with using raw or kernel-smoothed data. A sympathetic reader would care because the preprocessing step is cheap, independent of the learning model, and addresses a gap between DIC measurement and physics-informed machine learning.","feed_headline":"Positive-strain smoothing lifts DIC-based material modeling accuracy","feed_subtitle":"Better strain fields improve learned material models from two real biaxial experiments.","key_machinery":"The central object is the PGS optimization problem: reproduce the displacement field as a linear combination of reproducing-kernel (RK) basis functions, then choose the RK coefficients to minimize a hybrid loss, $\\text{loss}_u(\\bar{u}_I) + \\beta \\, \\text{loss}_E(\\bar{u}_I)$. Here $\\text{loss}_u$ measures agreement with the DIC measurements and $\\text{loss}_E = \\sum_J \\bigl|\\mathrm{ReLU}(-E_{11}[\\mathbf{u}](x_J)) + \\mathrm{ReLU}(-E_{22}[\\mathbf{u}](x_J))\\bigr|^2$ penalizes negative Green-Lagrange normal strains; the penalty weight $\\beta$ is rescaled by the initial ratio of the two losses so its value is length-scale independent. The RK basis acts as a low-pass filter for random sensor noise, while the strain-positivity term is the physics-informed component that suppresses smooth, non-physical compressive artifacts from operating errors. In the anisotropic valve dataset the same machinery is used protocol-by-protocol: only $E_{11}$ is penalized for stretch ratios $F_x:F_y = 1:1,\\,1:0.75,\\,1:0.5,\\,1:0.25$ and only $E_{22}$ for the three reversed ratios. The cleaned fields then feed a peridynamic neural operator that learns the nonlocal constitutive map.","core_discovery":"On the paper's own terms, the central discovery is that adding a physics-based loss term that penalizes negative normal strains while reconstructing displacements from DIC measurements removes non-physical compressive patterns, and that training a peridynamic neural operator on these cleaned fields reduces errors in the learned constitutive law. In the glove dataset the test relative $L^2$ displacement error drops from 10.01% for the rough DIC data and 9.14% for RK smoothing to 7.35% for PGS; in the heart-valve dataset it drops from 6.21% to 5.11%, and on the challenging subset of originally non-physical samples it drops from 17.85% to 14.68%. The paper also reports that models trained on PGS fields generalize better to unseen loadings and to different mesh resolutions, and that the learned kernel functions reveal interpretable isotropic or anisotropic, heterogeneous microstructure.","pith_inferences":["A natural next test is to apply the same positivity-constrained smoothing to synthetic biaxial datasets with known ground-truth strain fields, separating how much of the gain comes from removing noise versus from biasing labels toward positive strain.","The protocol-specific sign rule (penalize $E_{11}$ for one group of stretch ratios and $E_{22}$ for the other) suggests the method can be extended to more general loading protocols, but the sign selection would need to be automated or replaced by inequality constraints informed by an estimated Poisson effect.","Because the penalty weight $\\tilde{\\beta}$ controls a fidelity-versus-physicality tradeoff, the PGS output is effectively a one-parameter family of datasets; scanning $\\tilde{\\beta}$ could be used for uncertainty quantification in the downstream constitutive operator.","The same physics-based-loss-during-smoothing idea could transfer to other imaging modalities or to smoothing bases other than reproducing kernels, so long as the constraint is a trustworthy physical invariant."],"forward_implications":["Because the positivity penalty is applied during data preprocessing rather than inside the learning model, PGS can be dropped into existing DIC-to-material-modeling pipelines without changing the downstream architecture.","On the glove dataset, PGS lifts the benefit of smoothing beyond plain RK smoothing: the test error improves by roughly 20% relative to the smoothed baseline and about 27% relative to the rough data.","On the heart-valve dataset, the largest gain appears on the STest subset of originally non-physical samples, where the learned operator trained on PGS data reduces error from 17.85% to 14.68%.","Models trained on PGS-cleaned fields are less sensitive to test-resolution changes, retaining much of their accuracy on 16x16 and 31x31 grids.","The learned peridynamic kernels remain interpretable: isotropic on the glove and anisotropic with spatially varying fiber orientation on the valve leaflet."],"supporting_citations":[{"why":"Defines digital image correlation as the measurement technique whose noisy outputs PGS processes.","marker":"[2]"},{"why":"Supplies the reproducing-kernel basis functions used as the smoothing filter in PGS.","marker":"[8]"},{"why":"Provides the heterogeneous peridynamic neural operator used downstream to learn constitutive laws from the smoothed fields.","marker":"[5]"},{"why":"Prior work showing that RK-smoothed DIC data reduces generalization error in physics model learning, the baseline PGS builds on.","marker":"[21]"},{"why":"Represents prior art in regularized strain-field denoising for DIC that PGS extends with a physics constraint.","marker":"[7]"},{"why":"Documents that physical compressive strains can occur under biaxial tension via Poisson contraction, motivating the protocol-specific sign rule.","marker":"[22]"},{"why":"A second experimental study of biaxial tissue mechanics cited to justify that physical compressive strains can occur under unequal tensions.","marker":"[23]"},{"why":"Introduces peridynamic neural operators, the learning architecture used to evaluate the PGS-cleaned datasets.","marker":"[24]"}],"fun_headline_variants":["Physics-guided DIC smoothing improves material model accuracy","Positive-strain constraint sharpens DIC-based material learning","Cleaned DIC fields reduce errors in learned material models","DIC smoothing via positive strains sharpens model predictions","Physics-constrained DIC data yields better material models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a negative normal strain along the direction of the dominant biaxial stretch is always a non-physical artifact in these experiments, so penalizing it cannot remove real deformation.","fun_headline_variants_meta":{"raw":{"variants":["Physics-guided DIC smoothing improves material model accuracy","Positive-strain constraint sharpens DIC-based material learning","Cleaned DIC fields reduce errors in learned material models","DIC smoothing via positive strains sharpens model predictions","Physics-constrained DIC data yields better material models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000814,"raw_usage":{"total_tokens":3520,"prompt_tokens":850,"completion_tokens":2670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":2593}},"tokens_in":466,"tokens_out":2670,"duration_ms":17245,"temperature":1.0,"reasoning_tokens":2593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:25:19.836302+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PGS on a synthetic biaxial dataset whose ground-truth deformation is known and includes genuine compressive strains (e.g., from Poisson contraction under unequal tension), then compare the smoothed fields with the ground truth. If PGS systematically erases the true compressive regions or worsens the downstream learned operator compared with unconstrained smoothing, the sign rule is removing real physics rather than artifacts.","supporting_citations":[{"cited_title":"Digital image cor- relation,","cited_arxiv_id":null,"evidence_quote":"Defines digital image correlation as the measurement technique whose noisy outputs PGS processes."},{"cited_title":"Filters, reproducing kernel, and adaptive meshfree method,","cited_arxiv_id":null,"evidence_quote":"Supplies the reproducing-kernel basis functions used as the smoothing filter in PGS."},{"cited_title":"Heterogeneous Peridynamic Neural Operators: Discover Biotissue Constitutive Law and Microstructure From Digital Image Correlation Measurements","cited_arxiv_id":"2403.18597","evidence_quote":"Provides the heterogeneous peridynamic neural operator used downstream to learn constitutive laws from the smoothed fields."},{"cited_title":"Learningdeepimplicitfourierneural operators (IFNOs) with applications to heterogeneous materialmodeling,","cited_arxiv_id":null,"evidence_quote":"Prior work showing that RK-smoothed DIC data reduces generalization error in physics model learning, the baseline PGS builds on."},{"cited_title":"Strain field de- noisingfordigitalimagecorrelationusingaregularized cost-function,","cited_arxiv_id":null,"evidence_quote":"Represents prior art in regularized strain-field denoising for DIC that PGS extends with a physics constraint."},{"cited_title":"Biaxial me- chanical properties of the natural and glutaraldehyde treatedaorticvalvecusp—parti: experimentalresults,","cited_arxiv_id":null,"evidence_quote":"Documents that physical compressive strains can occur under biaxial tension via Poisson contraction, motivating the protocol-specific sign rule."},{"cited_title":"Mechanical characterisation of porcine rectus sheath under uniaxial and biaxial tension,","cited_arxiv_id":null,"evidence_quote":"A second experimental study of biaxial tissue mechanics cited to justify that physical compressive strains can occur under unequal tensions."},{"cited_title":"Peridynamic neural operators: A data-driven nonlocal constitutive model for complex material responses,","cited_arxiv_id":null,"evidence_quote":"Introduces peridynamic neural operators, the learning architecture used to evaluate the PGS-cleaned datasets."}],"review_version":1}